home / lab

The lab.

22 entries in five chapters. Each one says what state it's in, because "project" can mean anything from a running service to a notebook that ate a thousand dollars.

01 chapter

Agent systems and tooling

Harnesses, sandboxes, memory, simulation, and the frameworks I built before the popular ones existed.

an isometric arcade cabinet with a glowing screen, joystick, and two buttons, a single power cable trailing away

Arcade cabinet as a smart-home device

2025 – 2026Simulation-complete and deploy-ready. Hardware bring-up still open.

A living-room arcade cabinet where everything except running the retro games is offloaded to a configurable container on the home server. The cabinet itself does one job. Configuration, library, themes, brightness, and repair all happen over the network, so the device is set up and changed with almost no input on the machine itself. It is a smart-home device that happens to play games.

The cabinet is an MCP server: Joy and Home Assistant can list games, launch them, change themes, and dim the screensaver over the network
Repair agents run on the cabinet and are exposed over the network as a first-class feature; a broken config or a stale emulator gets diagnosed and fixed from the server, not with a keyboard on the floor
A config-as-code repo re-renders the emulator configs before every launch so nothing drifts
The remote library pushes games over a tunnel with checksums and resumable transfers
Joystick plus two buttons must reach every screen; anything else is filed as a theme bug
Prime directive: emulation keeps working with the network cable cut. Chaos soak of eight hours of injected partitions and daemon kills, zero violations

The cabinet is the thinnest possible client. Everything worth changing lives where an agent can reach it.

Python
Linux
MCP
WebSocket
RetroArch
Home Assistant
Toxiproxy

Arcade cabinet, in twenty-five seconds

An arcade cabinet that's really a smart-home device: favorites first, joystick plus two buttons, and every setting, game, theme, and repair handled from the server by an agent you talk to in plain language. Cut the cable and the games still run.

Download All thirteen films →

an isometric glass cube containing a single gear, with a plug socket on its side and a key

Splice

September 2025Installable package with CLI and docs. Built in two days.

Session-scoped LLM code execution. Define an edit point as a YAML contract with a signature, types, examples, and guardrails. A model writes exactly one function to fit. Splice runs it in a sandbox, checks the examples, caches the artifact, and re-executes without another model call. One LLM call, thousands of executions.

Child-process sandbox with filtered builtins, an import allowlist, and filesystem, network, and subprocess gates that default to off
Cache key is the whole determinism story: session, function, prompt hash, spec hash, context hash, config hash
Example functions with the wrong signature are rejected at load time, before a model is ever called
Auto-repair resends the same bundle with the exception and a minimal traceback, under a repair budget
Dependency policy off, if-present, or allow, with an ephemeral per-session virtualenv
40 tests across 16 files, one per build phase

AI-authored logic as a dependency-injected strategy object, with a contract, a sandbox, and a cache key.

Python
YAML
SQLite
mypy

Splice, in twenty-five seconds

Splice: write a YAML contract, let a model write exactly one function, run it in a sandbox, check the examples, cache it. One LLM call, thousands of executions.

Download All thirteen films →

VibeTest

December 2025Working local CLI with unit tests. Private, no public link.

A persona-simulation engine for testing messaging before it goes out. A blueprint DSL defines demographic distributions with weighted choices, normal curves, and conditional logic. Deterministic math builds persona skeletons with guaranteed statistical properties. A model adds grounded narrative. Concurrent interviews with A/B message variants are synthesized into consensus, variance, and outliers.

A "Mom Test" pass rewrites leading questions before they're asked
Persona demographics come from math, not the model, so a run with a 60/40 split is a 60/40 split
A/B variants are interviewed concurrently and scored on consensus, variance, and outliers
Its own output says the result is a hypothesis, not a finding

Synthetic personas generate hypotheses. The tool says so, in its own output.

Python
OpenRouter
asyncio

VibeTest, in twenty-five seconds

VibeTest: describe your audience as a blueprint, let math build fifty personas with the exact split you asked for, run twenty interviews at once with the leading questions rewritten, and get consensus, variance, and outliers, labeled as a hypothesis, not a finding.

Download All thirteen films →

an isometric handheld voice recorder with waveform ribbons entering a small crystal cube

Cortex

February 2026Working prototype. Offline since spring 2026.

Always-on voice to structured memory. A handheld recorder captures the day. Voice-activity detection segments audio locally. A cheap model triages each segment against a JSON schema: skip it, keep the on-device transcript, or send it to cloud transcription. Transcripts are diarized, then a twelve-category extraction pulls tasks, commitments, decisions, and people into DuckDB with embeddings, ready for the task board.

First real load: about 1.1 GB from a handheld recorder, 180 extractions, 18 direct commands and 162 ambient insights
Five parallel transcription workers turn a 540 MB recording around in 21 minutes
Twelve extraction categories: tasks, commitments, decisions, people, places, and more, each with a source segment
About three to four dollars a day to run at that volume
Android and Mac clients offload the recorder; a FastAPI service accepts uploads and signals when a day is ready

The day gets recorded once and comes back as a task list I didn't have to write.

Python
DuckDB
FastAPI
Silero VAD

Cortex, in twenty-five seconds

Cortex: wear a recorder all day, let VAD cut it into segments, let a cheap model decide which ones deserve cloud transcription, extract twelve categories with source timestamps, and let Joy file the tasks.

Download All thirteen films →

GBrain and the Living Atlas

2026Operating, with tests for HTTP and MCP contracts, graph, temporary access, and deployment

The pgvector-backed knowledge base behind Joy and the navigable graph surface over it. At last read it held 1,393 compiled pages and about 6,500 source chunks in a 465 MB database, built from massive initial ingestions of years of daily notes, meeting transcripts, message history, people records, documents, and books, and it grows every night. The Atlas adds role-scoped temporary access so an agent or a person can be handed a bounded window into memory without a permanent credential.

I wanted to be able to look at what the agents remember, and to lend that memory out safely.

Python
PostgreSQL
pgvector
MCP

GBrain and the Living Atlas, in twenty-five seconds

GBrain: the memory behind Joy. About 1,400 compiled pages and 6,800 source chunks on pgvector, every fact cited to a source with a time boundary, nightly updates that say what was learned, and a Living Atlas that can lend a bounded window into memory without handing out a credential.

Download All thirteen films →

thumbnail from the original resume site for the first-generation Cyrus

Cyrus, first generation

2023Superseded by Cyrus-v2

A 2023 project built on the original release of LangChain, with large parts of its underlying mechanisms reimplemented from scratch: an event-driven multi-agent framework with an asyncio state-machine flow engine, native tool calling, model routing, Bedrock and Perplexity agents, a VS Code extension, and a Streamlit UI. The first working draft landed the week AutoGen launched.

I wanted reproducible agent workflows before the frameworks did.

Python
LangChain
asyncio
TinyVDB

Cyrus, first generation, in twenty-five seconds

Cyrus, first generation (2023): an event-driven multi-agent framework on the original LangChain, with a state-machine flow engine, one tool-calling contract across GPT-4, Gemini, Claude, and Bedrock, cost on every call, a shared workspace, a VS Code extension, and a vector DB with no server. It worked the week AutoGen launched.

Download All thirteen films →

TinyVDB

2024Used inside Cyrus and other projects

A vector database that runs inside your application with built-in embeddings, cosine and Euclidean search, metadata filtering, and disk persistence. No server to run.

Most projects need a vector index. Few need a vector service.

Python

Spec-driven data-migration agents

2025Innovation-sprint prototype at my employer

A fork of the Gemini CLI with spec-driven development baked in. A project type is defined in JSON: stack, scaffolding, guardrails, and how a feature should be created. Sub-agents follow the spec, and a search sub-agent surfaces prior patterns from mock internal tools.

Repeatable, reviewable feature skeletons with the right files in the right places, ready for a human to finish.

TypeScript
Gemini CLI

An afternoon with an AWS product team

2026Delivered. One session.

Consulted on a roughly $100 million enterprise deal. An AWS product team was positioning what became Amazon Quick Suite against n8n and Make, and they needed someone who had run n8n hard enough to know where it breaks at enterprise scale: who can see the data, who can audit a run, what happens to credentials, and what happens when the vendor pivots. I reviewed the product, walked through each of those failure modes, and told them which ones their design already answered. Independent advisory, not a partnership or an endorsement.

Explaining a technical choice to people who won't read the code is most of the job.

Amazon Quick Suite
Bedrock Flows
n8n
Make

02 chapter

Models and data

Training runs, classifiers, retrieval before the acronym, and one thousand-dollar mistake.

an isometric GPU card with a small crack and a spilled coin jar, one tiny checkpoint flag missing from a flag stand (no text)

The AI image detector that forgot to checkpoint

2023Training run lost before checkpoint. Kept for the lesson.

I wanted a diffusion-image detector that generalized across models instead of one classifier per architecture. I assembled 4.1 million images from fifteen sources, trained a class-attention transformer on SageMaker, and lost the whole run when the time cap hit before a checkpoint had been written.

4.1 million images from fifteen sources, 748 GB raw, 150 GB after processing
36-hour training run, about a thousand dollars of credits
Zero checkpoints saved, because I misconfigured the checkpoint path
I now design the checkpoint before the model

Everyone has a story like this. Mine has a receipt.

PyTorch
AWS SageMaker

DeepQuote

2019 – 2020Complete

GPT-2 fine-tuning on a narrow conversational corpus to reproduce individual writing styles, across the 124M, 335M, and 1.5B models. Discovered that 9.6% of generated quotes were verbatim training data. My first real lesson in overfitting.

Where "building with AI since 2019" actually starts.

Python
GPT-2
Google Colab

PDF document QA, August 2022

2022Complete

Heuristic summarization to shrink a document into a context window, keyword and synonym weighting from the question, then OPT-1.3B and later GPT-3 answering in a command-line chat. Built to read hundreds of articles for a business plan, three months before ChatGPT.

Retrieval-augmented generation before the acronym.

Python
OPT-1.3B
GPT-3

FIRESight

2024Prototype

Adaptive Monte Carlo financial-independence simulation with over 150 parameters, where Bedrock-hosted agents add parameters, request inputs, and generate new simulation logic in response to "what if" questions.

An early attempt at agents that modify the model rather than just run it.

Python
AWS Bedrock

Cat facial recognition

2023Research project, complete

Non-invasive biometrics for identifying missing cats from video. OpenCV cascade extraction, a CNN and an SVM on a pre-trained backbone, 98% and 100% on held-out frames from twelve cats.

Small dataset, careful evaluation, a real animal-welfare use.

Python
OpenCV
PyTorch

AWS DeepRacer

2024Game day

Reinforcement learning on a real track: state-dependent rewards, a hand-crafted discrete action space, a penalty on steps per progress, and racing-line waypoints as a supervised hint.

Reward design is system design.

Python
AWS DeepRacer

SewingPatternSearch

2025Working Streamlit tool

Text and visual search over a library of sewing-pattern PDFs. Multimodal retrieval for a domain that has none.

Built for a real person with a real shelf of PDFs.

Python
Streamlit

03 chapter

Hardware, home, and a book

Things with uptime, things on a convention floor, and one argument in progress.

an isometric helmet with an LED dot-matrix face, sound waves entering one side

Wake-word training and the LED mask

2025Worked on a convention floor. Windows, local-only.

A helmet with an LED dot-matrix face and a webcam inside. It reads the other person's expression and emotes back, and it reacts to keywords through a wake-word model trained entirely on synthetic speech. Everything runs offline on a handheld PC.

The mask's detector is a modified OpenWakeWord pipeline trained on synthetic speech, with a crowd-noise augmentation pass because convention floors are loud
My reusable wake-word demo synthesizes training data with three TTS voices, three phrasings, and three speaking rates, plus thirteen deliberate near-miss negatives
That demo adds gain jitter, Gaussian noise, and time shifts, trains a tiny MLP over streaming embeddings, and exports to ONNX for CPU
Hundreds of expressions pre-generated as images and selected at runtime from emotion and keyword state
Face tracking keeps eye contact aligned as the wearer moves

screenshot of the LED mask's telemetry, valence and arousal plots and the avatar face window

Robustness was engineered from augmentation rather than collected from recordings.

Python
OpenWakeWord
ONNX
InsightFace
OpenCV

Wake-word training and the LED mask, in twenty-five seconds

A helmet with an LED face: it reads your expression and emotes back, and it wakes on 'Hey Joy' with a model trained on nothing but synthetic speech: three voices, three phrasings, three rates, thirteen near-miss negatives, noise and jitter on top. Offline, on a handheld PC, on a convention floor.

Download All thirteen films →

Media operations platform

2022 – presentOperating

Request-to-availability automation for a large self-hosted library: intake, search, fetch, import and repair passes, quality profiles, subtitle handling, and a parity check that confirms a fulfilled request is actually playable. The fetch tier runs behind Gluetun, a containerized VPN client that owns the network namespace for three services, with a kill switch and a small port-forwarder that reads Gluetun's forwarded port and updates the client, so nothing leaks and nothing stalls when the tunnel rotates.

Idempotent fulfillment across a network boundary you don't control is the same problem shape as any third-party data pipeline.

Docker
Gluetun
Unraid
Python

an isometric home server tower with many small container boxes stacked inside like drawers, one drawer open

Lampyr

2022 – presentOperating. About fifty containers across nine Compose stacks.

The Unraid host under everything: agent runtimes, browser infrastructure, databases and queues, productivity services, media, and household automation. I ran two Kubernetes clusters here once and simplified on purpose.

You learn what operational debt means by owing it to yourself.

Unraid
Docker Compose
PostgreSQL
Redis
RabbitMQ

Fabrication tools

2025Delivered CAD, layouts, and generators; some builds unconfirmed

A local-first Jenga-set generator with a Flask API and React front end that renders laser-ready layouts from folders on disk, plus parametric CAD, lithophane image processing, and a seven-foot 3D-printed relief.

Deterministic outputs, and local when privacy or payload size makes cloud the wrong choice.

Python
Pillow
Flask
React
Fusion 360

XKCD screensaver

2022Shipped in 2022. Still runs.

A Lambda that syncs the daily comic to a TV through Flickr. The smallest useful system on this page.

Small, scheduled, and four years old.

Python
AWS Lambda

A book on machine consciousness

2026 – in progressIn progress, with a friend

A collaborative research and writing project on subjective experience, world models, embodiment, global workspace theory, and interpretability. My position is provisional and leans toward possible machine experience without claiming current systems have it.

Because I think the question is real, and because writing is how I find out what I think.

Reading
arguing

+ systems

Four more systems with full routes.

Not on the home page, still with a diagram and a role.